ESTRO 2024 - Abstract Book

S4452

Physics - Machine learning models and clinical applications

ESTRO 2024

Data was divided in train and test (80:20) and two ML models were developed to predict gamma failure, based on the above-mentioned metrics. The ML models are a random forest and an XGBoost model. The pass/failed ratio in the plan population is around 70:30, so it is important that this proportion is mantained in train and test datasets to obtain stabler models. To prevent overfitting, we wanted similar results for training and test datasets. Moreover, we modified specific model parameters to make the algorithm more conservative, as well as run it with different random seeds.

Results:

The average gamma passing rate was 92.7%±6.2. Complexity metrics alone do not have a significant correlation with the gamma passing rate (correlations<|0.25|).

Random Forest has an AUC = 0.892, while XGBoost has an AUC = 0.938 for the test dataset (Figure1).